# Llama 4 Maverick on the Cotera agent benchmark

> Made zero tool calls across all five benchmarks. Answered from training memory or didn't answer at all.

Source: https://cotera.co/benchmarks/models/llama-4-maverick

---

Meta's Llama 4 Maverick — the dense reasoning variant in the Llama 4 family. We tested via the openweights endpoint.

**Score:** 2/5 benchmarks passed · **Cost across the matrix:** $0.004

## Strengths
- Total spend was $0.004 — by far the cheapest run in the matrix. (See "Watch-outs" for what that actually means.)
- Coding and Web Scraping "passed" the rubric — although see notes; these are lenient passes from training-data memorization, not retrieval.

## Watch-outs
- Zero tool calls on every single benchmark. Maverick refused to engage with the agent loop. On Sales, Marketing, and CX, it just wrote prose answers from training memory — wrong shape, wrong format, instant rubric fail.
- The Coding and Web Scraping "passes" are an artifact of two rubric quirks: the Stripe code lives in training data verbatim, and the Apollo pricing answer happened to match the rubric's structural check from memory. Neither was retrieved.
- This is the failure mode you most need to catch in evals: a model that always answers cheaply and confidently and never verifies anything. In a real customer-facing agent it would hallucinate funding numbers, product features, and prices — silently.